Found while building the qa-run roll-up (#9486 / PR #9616), which had to parse these titles for real.
The drift
.claude/skills/checklist-test/SKILL.md:126 documents the run-record title as five ·-separated fields with a standalone judged/total:
QA run · SELECTOR · JUDGED/TOTAL · FRAMEWORK-SHA8 · DATE
No record has ever been written that way. Measured over the complete live corpus on 2026-08-18 — all 23 qa-run issues, open and closed:
- 23 / 23 use five fields in which the judged/total is folded into a parenthetical on the selector, and a counts field the documented convention does not mention at all:
QA run · SELECTOR (JUDGED/TOTAL) · FRAMEWORK-SHA8 · DATE · COUNTS
The card for #9486 described a third shape (six fields, standalone judged/total and counts). Zero records follow that one either.
Why it matters
The documented line is the only written spec for a field that is now parsed by tooling. PR #9616 reads these titles to produce the QA roll-up, and had to be written against the measured shape rather than the documented one. The next person to write against SKILL.md will produce a parser that silently matches nothing — or, worse, one that mis-assigns fields by position and reports confident nonsense.
(FULL area) is also doing real damage as a phrasing: it declares no total, and at least one record titled (FULL area) judged 13 of its area's 33 items (#9351, records-forms). The roll-up now renders that as 13/33* by reading the item count from docs/qa/platform-checklist/areas/, but the title alone reads as complete coverage.
Suggested resolution — needs a maintainer, not a dev
.claude/skills/** is a governed, human-merge-only surface, so this is filed rather than fixed.
Two directions, and they are not equivalent:
- Update SKILL.md to document what runs actually write. Cheap, immediate, keeps every existing record valid. Leaves five phrasings of the same field as the permanent contract.
- Tighten the convention and have runs emit one shape. A parser could then be strict rather than tolerant, which is the contract-first direction. Costs a migration decision for the 23 existing records — most cleanly, "new records only", since the roll-up already reports what it cannot parse.
Recommendation: 2, with the roll-up's "Not parsed" section as the enforcement surface — it already prints any record it cannot read, so a tightened convention becomes visible the first time someone deviates instead of silently degrading. 1 alone leaves the tolerant-consumer shape that hides authoring errors.
Filed unassigned for triage.
Generated by Claude Code
Found while building the qa-run roll-up (#9486 / PR #9616), which had to parse these titles for real.
The drift
.claude/skills/checklist-test/SKILL.md:126documents the run-record title as five·-separated fields with a standalone judged/total:No record has ever been written that way. Measured over the complete live corpus on 2026-08-18 — all 23
qa-runissues, open and closed:(18/18),(77 items),(FULL area),(6/14 items consulted),(5 of 10 items)priority:P0)/ 8 NOT-RUNsegment and a trailing(11 not-run)parentheticalThe card for #9486 described a third shape (six fields, standalone judged/total and counts). Zero records follow that one either.
Why it matters
The documented line is the only written spec for a field that is now parsed by tooling. PR #9616 reads these titles to produce the QA roll-up, and had to be written against the measured shape rather than the documented one. The next person to write against SKILL.md will produce a parser that silently matches nothing — or, worse, one that mis-assigns fields by position and reports confident nonsense.
(FULL area)is also doing real damage as a phrasing: it declares no total, and at least one record titled(FULL area)judged 13 of its area's 33 items (#9351,records-forms). The roll-up now renders that as13/33*by reading the item count fromdocs/qa/platform-checklist/areas/, but the title alone reads as complete coverage.Suggested resolution — needs a maintainer, not a dev
.claude/skills/**is a governed, human-merge-only surface, so this is filed rather than fixed.Two directions, and they are not equivalent:
Recommendation: 2, with the roll-up's "Not parsed" section as the enforcement surface — it already prints any record it cannot read, so a tightened convention becomes visible the first time someone deviates instead of silently degrading. 1 alone leaves the tolerant-consumer shape that hides authoring errors.
Filed unassigned for triage.
Generated by Claude Code